GPU chess update, local memory....

[smatovic]

I have AMD Radeon 6900 series with 2 GB GPU RAM.

Not sure if this device supports the relevant atomic features.

But if OpenCL is properly installed then the device should be checked by --guessconfig
and print Error lines if features are not present.

You can add the option -l to the command line call to activate logging,
errors are printed into the file zeta.log .

Maybe you can post or pm me the zeta.log file....

--
Srdja

[smatovic]

...ah, silly me,

i forgot that gpus, compared to native x86-64 cpus, need multiple cycles for 64 bit integer operations, so my numbers look not that bad as i thought.

A switch to an 32 bit float or 24 bit based design could bring the desired x4 speedup....

--
Srdja

[noobpwnftw]

GPU is not optimized for control logic, which is used extensively in chess engines. Even if you do get the same or better performance with NPS, it is still not efficient when pruned search is introduced. You would either waste too much on cut-offs(lazy) or waiting for the slowest child(ybwc) or just global hash probe warps.

[smatovic]

I agree, it is one task to get reasonable nps throughput and another one to implement efficient parallel search.

[smatovic]

How do I tell it to examine the GPU?

I've added command line options to show all OpenCL devices,

"zeta.exe -dl",

and to select an specific device for guessconfig,

"zeta.exe -p i -d j --guessconfigx",

see readme file for further details.

https://github.com/smatovic/Zeta/tree/dev

..will update the binary in a minute...

--
Srdja